GPU databases are coming of age
GPUs are that obscure object of desire right now. Originally created to provide better performance for gamers, now everyone from crypto miners to deep-learning experts wants a piece of them.
Increased demand for cryptocurrency mining and competition for memory modules has created a perfect storm driving GPU prices to skyrocket. That’s bad news for users, but great for GPU manufacturers like Nvidia.
We have elaborated onZDNet on how Nvidia has gone from gamer’s delight to AI powerhouse. Just looking at the agenda at GTC, Nvidia’s GPU Technology Conference taking place this week in Silicon Valley, confirms this transition.
But why should you care about GPUs if you’re not into AI, gaming, or crypto? Because GPUs can also accelerate your databases, and there’s not a single organization today not using one.
GPUs greatly accelerate operations that can be parallelized. This approach has been in use in masively parallel architectures such as Hadoop or Spark for a while now. The idea is to combine an array of database instances, each on a separate server, and then to use a master node that delegates subqueries to each one.
The individual servers execute their subqueries in parallel, get the result sets back to the master node, which combines them and sends a single one back to the client. GPUs make the same divide-and-conquer approach possible within individual servers, with CPUs taking the role of the master node.
So, can you just add some GPUs to the server hosting your database and expect to see a massive improvement in performance? Not so fast. First of all, not all database operations are parallelizable, and for the ones that are not, adding GPUs won’t make a difference.
But even for operations that can be parallelized, databases have to be designed and implemented in a way that enables them to take advantage of the GPU architecture. In other words, it takes a special type of database to be able to capitalize on GPUs.
This is the premise on which GPU databases such as SQream were built on. SQream was born when entrepreneur and investor Amit Gai met programmer and algorithm builder Kostya Varakin in 2010. They joined forces and started working on implementing Varakin’s ideas on how to work with GPUs in a database.
A few years, patents, and funding rounds later, SQream is run by Gal and VP R&D Razi Shoshani. Varakin has set off to new adventures, but the company has raised $15 million to date and has more than 50 employees and is expected to grow to around 75 employees during 2018.
Today, at GTC, SQream is announcing the latest in a series of partnerships, and we took the opportunity to connect and discuss. Before we delve on the specifics of the partnership, however, let’s take a step back to quickly review this space.
Apart from SQream, other GPU database players include Blazegraph, BlazingDB, Brytlyt, Kinetica, MapD, and PG-Strom. SQream positions itself in the analytics database market and differentiates by focusing on big workloads. According to David Leichner, CMO SQream:
“There are several vendors who develop and sell GPU databases. The ones you probably hear the most about are Kinetica and MapD. While SQream is often coupled with these two vendors in roundups about GPU databases, there is one main and several sub differentiators between the solutions. First and foremost, Kinetica and MapD use in-memory storage. While this enables them to provide extremely fast analysis of up to say 5TB or 10TB of data, they are limited in scalability due to architecture as well as cost. You rarely see either of these talking about data stores of 20TB and up. SQream, on the other hand, is built for massive data stores.
